Interview questions for AI fluency: 12 questions to assess AI skills, with a 1–4 scale

To assess AI skills in an interview, ask candidates to walk through a real task they did with an AI tool, how they checked what it produced, and when they chose not to use it, then score the answer against anchors written in advance. The questions below are grouped under the four competencies in the AI Fluency Framework published in 2025 by Rick Dakan, Joseph Feller and Anthropic: delegation, description, discernment and diligence. Each of the twelve questions comes with what a strong answer shows, a follow-up probe and 1–4 anchors, and there is a ten-minute output-review exercise that every candidate can do regardless of which tools their last employer allowed.

This page treats AI use as a job skill. It is not about catching candidates who use AI to answer your questions; for that, see interview questions that expose AI-assisted answers and how to write a policy on candidates using AI in interviews.

What does AI fluency mean in a hiring context?

The framework defines AI fluency as interacting with AI systems in ways that are effective, efficient, ethical and safe, and splits it into four competencies. Translated into what you can hear in an interview:

AI fluency is not knowing the newest product names, clever prompt tricks or enthusiasm for automation. A candidate who uses one tool well and checks everything it produces is more fluent than one who lists ten tools and trusts all of them. It sits close to data literacy and judgment, and the discernment questions overlap with critical thinking.

Delegation: does the candidate know when to use AI?

1. "Walk me through one task in your current job that you now do with an AI tool, from start to finish."
   • Strong answer shows: a specific, recurring task; what the tool does and what the candidate still does; what changed in time or quality.
   • Probe: "Which part did you keep for yourself, and why that part?"
   • 4: A workflow described end to end, a clear split of responsibilities, a measured change, and an adjustment made after something went wrong.
   • 3: A specific task with a clear split and a plausible benefit.
   • 2: General use ("I use it for emails") with no split or result.
   • 1: No example, or only tool names.
2. "Tell me about a task where you decided not to use AI."
   • Strong answer shows: a named reason, such as confidential data, accuracy that had to be exact, a policy, a relationship that needed a personal touch, or simply that it was faster by hand.
   • Probe: "What would have had to be true for you to use it?"
   • 4: A real decision with a clear reason, and a rule they now apply to similar tasks.
   • 3: A real decision with a sensible reason.
   • 2: A vague reason ("it didn't feel right"), or the only limit they name is the tool's quality.
   • 1: Uses AI for everything and cannot name a limit, or avoids it entirely without being able to say why.
3. "Describe a time an AI tool saved you less time than you expected, or cost you time."
   • Strong answer shows: honest accounting that includes checking and rework, and a changed approach.
   • Probe: "How did you know how long it actually took?"
   • 4: Counted the full cost including review, changed or dropped the workflow, and can say where it does pay off.
   • 3: A specific example and a sensible conclusion.
   • 2: Says it always saves time; no example of it failing.
   • 1: No example.

Description: can the candidate brief an AI tool well?

4. "Here is a task from this job: [a real task, such as drafting a weekly client update from three sets of notes]. How would you brief an AI tool to help with it?"
   • Strong answer shows: the goal, the audience, the source material, constraints, the format, and an example of good output.
   • Probe: "The first draft comes back generic. What do you change?"
   • 4: Covers context, audience, constraints and format unprompted, and improves the brief rather than retrying the same request.
   • 3: Covers most of these; adds the rest when probed.
   • 2: A one-line request; the fix is "ask it again".
   • 1: Cannot describe how they would start.
5. "Tell me about a time you got a poor result from an AI tool and improved it."
   • Strong answer shows: a diagnosis of why the result was poor, such as missing context or an ambiguous goal, and the change that fixed it.
   • Probe: "What exactly did you change in what you gave it?"
   • 4: Clear diagnosis, a specific change, a better result, and the lesson reused since.
   • 3: A specific fix that worked.
   • 2: Kept regenerating until something acceptable came out.
   • 1: No example.
6. "How do you make something that worked once repeatable, for you or your team?"
   • Strong answer shows: saved instructions or templates, notes on what to check, and sharing with colleagues.
   • Probe: "Has anyone else used it? What did they change?"
   • 4: Built something colleagues use, and revised it after it failed for someone.
   • 3: Has a personal template or saved brief they reuse.
   • 2: Starts from scratch every time.
   • 1: No repeatable use.

Discernment: does the candidate check the output?

7. "Tell me about a time an AI tool gave you something wrong that looked right."
   • Strong answer shows: the specific error, such as an invented source, a wrong figure or an outdated rule, how they caught it, and the check they now run.
   • Probe: "Would you have caught it if you had been in a hurry?"
   • 4: A specific error caught by their own check, a permanent change to how they verify, and an honest view of when they might still miss one.
   • 3: A specific error and how they caught it.
   • 2: Knows tools make mistakes but has no example of catching one.
   • 1: Has never noticed an error, or someone else caught it after it was sent.
8. "What do you check before AI-assisted work goes to a client or your manager?"
   • Strong answer shows: checks scaled to the risk: names, numbers and claims against the source; the whole text read before sending; more care for anything external or legal.
   • Probe: "Which kinds of output do you check line by line, and which do you skim?"
   • 4: A clear, risk-based routine with an example of it changing what was sent.
   • 3: A sensible routine, stated specifically.
   • 2: "I read it over"; no difference between low- and high-risk work.
   • 1: Sends output unchecked.
9. "Tell me about a time you disagreed with what an AI tool recommended."
   • Strong answer shows: what it missed that the candidate knew, how they weighed it, and the outcome.
   • Probe: "Looking back, was the tool right about anything?"
   • 4: Weighed it fairly, explained the gap in its context, decided, and checked the result.
   • 3: A real disagreement with a clear reason.
   • 2: Always accepts or always rejects its suggestions.
   • 1: No example.

Diligence: does the candidate handle data and ownership responsibly?

10. "What information would you never put into an AI tool at work, and why?"
    • Strong answer shows: personal data, client confidential material, credentials and anything their employer's policy excludes; awareness that tools differ in how they treat data; asking when unsure.
    • Probe: "What did you do the last time you were not sure?" Do not ask them to describe the confidential material itself.
    • 4: Clear categories, knows the policy, has checked a tool's data terms or asked, and has an example of holding back.
    • 3: Names the main categories and the reason.
    • 2: Vague ("nothing sensitive") with no example.
    • 1: Sees no issue with pasting anything in.
11. "When do you tell people that AI was involved in work you hand them?"
    • Strong answer shows: a rule tied to context, such as client deliverables, published work or anything the recipient will rely on, and ownership of the content either way.
    • Probe: "Was there a time you chose not to say? Why?"
    • 4: A consistent, reasoned rule, applied in a real example, consistent with the employer's policy.
    • 3: A sensible view with an example.
    • 2: Has not thought about it.
    • 1: Hides it where disclosure was expected.
12. "Tell me about a mistake in AI-assisted work that reached someone else."
    • Strong answer shows: the mistake owned as their own, told to the person affected, fixed, and a check added.
    • Probe: "What did you say to the person affected?"
    • 4: Owned it fully, told people promptly, fixed it, changed the process.
    • 3: Owned and fixed it.
    • 2: "The AI got it wrong"; their part is unclear.
    • 1: Blames the tool or did not tell anyone.

A ten-minute output-review exercise

Give every candidate the same short source and an AI-drafted message, both invented for the exercise, and ask: "An AI tool drafted this from the notes. What would you change before sending it?" No tool is needed, so candidates whose last employer banned AI are not disadvantaged.

Before the first interview, write down what a strong review catches: the discount was requested, not agreed, and depends on a 24-month term; current volume is 1,200, not 1,500; the "standard 12-month term" is not in the notes; and legal has not approved anything, so "this week" is a promise nobody made. A strong candidate finds all four, says they would check the draft against the notes rather than trust its tone, and rewrites the message accurately. Score it on the discernment anchors above.

An overall 1–4 scale for AI fluency

Set the level each role needs before interviewing; a 3 may be plenty for most roles. How to write and calibrate anchors like these is in behavioral anchors for interview scores, and for follow-ups beyond the ones listed, see probing questions for interviewers.

How do you keep AI fluency questions fair?

• Make it job-related. Write down which tasks in the role involve AI tools. If none do, leave the competency out.
• Plan for unequal access. Some employers ban AI tools and some candidates cannot pay for them. Decide in advance that examples from study, volunteering or personal projects count, and use the exercise for everyone.
• Score judgment, not products. Never score which tool someone uses or whether they pay for it.
• Protect confidential information. Ask how they decided what to share, not what the confidential material was.
• Avoid proxies for age. Phrases such as "digital native" describe a generation, not a skill. The behaviors above can be observed in anyone; see age discrimination in hiring.
• Ask the same questions of everyone for the role, and score each answer before the debrief.

Common mistakes when interviewing for AI skills

• Rewarding enthusiasm. "I use AI for everything" is a risk signal until the candidate explains how they check it.
• Penalizing caution. A candidate who explains why they kept AI away from client data is showing diligence, not reluctance.
• Testing prompt tricks. Clever phrasing ages quickly; context, constraints and checking do not.
• Accepting time-saved claims at face value. Ask whether review and rework were counted.